Pneumonia is an inflammatory condition of the lung is affected primarily the small air sacs known as alveoli. Symptoms typically include some combination of productive or dry cough, chest pain, fever and difficulty breathing. The severity of the condition is variable. Pneumonia is usually caused by infection with viruses or bacteria. Pneumonia isn’t just one disease. It is a family of more than 200 different lung diseases. Statistics for Pneumonia disease detection in Egypt detected that Pneumonia deaths is about 13,393 with 2.50 % and Rate of 20.69 with a World Rank #100. According to the latest World Health Organization (WHO) data published in 2020 Lung Disease Deaths in Egypt reached 13,393 or 2.50% of total deaths. Our main goal is earlier detection of Pneumonia by predicting a patient’s severity of decline in lung function and based on Chest X-ray of their lungs. Using image processing and machine intelligence techniques to help Pneumonia impacted patients and produce a prediction with the images of Pneumonia.
CNN stands for Convolutional Neural Network which is a specialized neural network for processing data that has an input shape like a 2D matrix like images. CNN's are typically used for image detection and classification.
Transfer learning is a machine learning technique where a model trained on one task is re-purposed on a second related task.!pip install opencv-python
Requirement already satisfied: opencv-python in d:\anaconda\lib\site-packages (4.7.0.68) Requirement already satisfied: numpy>=1.17.0 in d:\anaconda\lib\site-packages (from opencv-python) (1.21.5)
# Making all necessary imports
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
import os
import glob
import json
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import tensorflow as tf
import cv2
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D,GlobalAveragePooling2D, Flatten, Dense, Dropout, BatchNormalization
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau
from tensorflow.keras.optimizers import Adam
import warnings
warnings.filterwarnings('ignore')
import logging
logger = tf.get_logger()
logger.setLevel(logging.ERROR)
print(tf.__version__)
2.4.1
def get_files(base_dir, target_dir):
count = 0
path = get_path(base_dir, target_dir)
for dirname, _, filenames in os.walk(path):
for filename in filenames:
count+=len(glob.glob(os.path.join(dirname, filename)))
return path, count
def get_path(base_dir, target_dir):
path = os.path.join(base_dir,target_dir)
return path
Dataset analysis is a necessary step before creating models
3 folders are provided to us:
Dataset/
test/
opacity/
normal/
train/
opacity/
normal/
val/
opacity/
normal/base_dir = 'D:/Machine Proj/Project/Dataset'
train_normal_dir = 'D:/Machine Proj/Project/Dataset/train/normal'
train_pneumonia_dir = 'D:/Machine Proj/Project/Dataset/train/opacity'
val_normal_dir = 'D:/Machine Proj/Project/Dataset/val/normal'
val_pneumonia_dir = 'D:/Machine Proj/Project/Dataset/val/opacity'
test_normal_dir = 'D:/Machine Proj/Project/Dataset/test/normal'
test_pneumonia_dir = 'D:/Machine Proj/Project/Dataset/test/opacity'
train_normal_path, train_normal_count = get_files(base_dir,train_normal_dir)
train_pneumonia_path, train_pneumonia_count = get_files(base_dir,train_pneumonia_dir)
val_normal_path, val_normal_count = get_files(base_dir,val_normal_dir)
val_pneumonia_path, val_pneumonia_count = get_files(base_dir,val_pneumonia_dir)
test_normal_path, test_normal_count = get_files(base_dir,test_normal_dir)
test_pneumonia_path, test_pneumonia_count = get_files(base_dir,test_pneumonia_dir)
print("No of Train Images: {}".format(train_normal_count + train_pneumonia_count))
print(" \u2022 No of Normal Images {}".format(train_normal_count))
print(" \u2022 No of Pneumonia Images {}".format(train_pneumonia_count))
print("No of Validation Images: {}".format(val_normal_count + val_pneumonia_count))
print(" \u2022 No of Normal Images {}".format(val_normal_count))
print(" \u2022 No of Pneumonia Images {}".format(val_pneumonia_count))
print("No of Test Images: {}".format(test_normal_count + test_pneumonia_count))
print(" \u2022 No of Normal Images {}".format(test_normal_count))
print(" \u2022 No of Pneumonia Images {}".format(test_pneumonia_count))
No of Train Images: 4192 • No of Normal Images 1082 • No of Pneumonia Images 3110 No of Validation Images: 1040 • No of Normal Images 267 • No of Pneumonia Images 773 No of Test Images: 624 • No of Normal Images 234 • No of Pneumonia Images 390
train_data = []
for filename in os.listdir(train_normal_path):
train_data.append((os.path.join(train_normal_path,filename),0))
for filename in os.listdir(train_pneumonia_path):
train_data.append((os.path.join(train_pneumonia_path,filename),1))
train_data = pd.DataFrame(train_data, columns=['image_path', 'label'], index=None)
train_data = train_data.sample(frac=1).reset_index(drop=True)
val_data = []
for filename in os.listdir(val_normal_path):
val_data.append((os.path.join(val_normal_path,filename),0))
for filename in os.listdir(val_pneumonia_path):
val_data.append((os.path.join(val_pneumonia_path,filename),1))
val_data = pd.DataFrame(val_data, columns=['image_path', 'label'], index=None)
test_data = []
for filename in os.listdir(test_normal_path):
test_data.append((os.path.join(test_normal_path,filename),0))
for filename in os.listdir(test_pneumonia_path):
test_data.append((os.path.join(test_pneumonia_path,filename),1))
test_data = pd.DataFrame(test_data, columns=['image_path', 'label'], index=None)
print("Train Data {}".format(train_data.shape))
print("Validation Data {}".format(val_data.shape))
print("Test Data {}".format(test_data.shape))
Train Data (4192, 2) Validation Data (1040, 2) Test Data (624, 2)
train_data
| image_path | label | |
|---|---|---|
| 0 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 1 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 2 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 3 | ../input/pneumonia-xray-images/train/normal/IM... | 0 |
| 4 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| ... | ... | ... |
| 4187 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 4188 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 4189 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
| 4190 | ../input/pneumonia-xray-images/train/normal/IM... | 0 |
| 4191 | ../input/pneumonia-xray-images/train/opacity/p... | 1 |
4192 rows × 2 columns
class_dict = {0:'Normal', 1:'Pneumonia'}
train_data['class_name'] = train_data.label.map(class_dict)
train_data['class_name'].value_counts().plot(kind='bar')
<AxesSubplot:>
for filepath in train_data.image_path:
image = cv2.imread(filepath)
image_size = image.shape
break
image_size
(1006, 1404, 3)
def visualize_img(images):
fig = plt.figure(figsize=(20, 15))
for i,path in enumerate(images):
fig.add_subplot(4, 4, i+1, xticks=[], yticks=[])
img = cv2.imread(path)
plt.imshow(img)
plt.title(train_data[train_data.image_path == path].class_name.values[0])
for i in range(2):
images = train_data[train_data.label == i].image_path
images = np.random.choice(images , 8)
visualize_img(images)
def plotImages(images_arr):
fig, axes = plt.subplots(1, 5, figsize=(20,20))
axes = axes.flatten()
for img, ax in zip(images_arr, axes):
ax.imshow(img)
plt.tight_layout()
plt.show()
BATCH_SIZE = 32
IMG_SHAPE = 224
train_image_gen = ImageDataGenerator(rescale=1./255,
width_shift_range=0.1,
height_shift_range=0.1,
brightness_range=[0.2,1.0],
zoom_range=0.2,
horizontal_flip=True,
fill_mode='nearest')
train_gen = train_image_gen.flow_from_dataframe(train_data,
x_col='image_path',
y_col='class_name',
class_mode='binary',
batch_size=BATCH_SIZE,
shuffle=True,
target_size=(IMG_SHAPE,IMG_SHAPE))
Found 4192 validated image filenames belonging to 2 classes.
augmented_images = [train_gen[0][0][2] for i in range(5)]
plotImages(augmented_images)
from sklearn.preprocessing import LabelBinarizer
from tensorflow.keras.utils import to_categorical
train_lb = to_categorical(train_data.label, dtype = int)
val_lb = to_categorical(val_data.label, dtype=int)
train_data = train_data.reset_index().drop(labels='index', axis=1)
y_train = pd.DataFrame(train_lb).add_prefix('label_')
val_data = val_data.reset_index().drop(labels='index', axis=1)
y_val = pd.DataFrame(val_lb).add_prefix('label_')
train_data = pd.concat([train_data, y_train], axis=1)
val_data = pd.concat([val_data, y_val], axis=1)
print("Training set has {} samples".format(train_data.shape[0]))
print("Validation set has {} samples".format(val_data.shape[0]))
Training set has 4192 samples Validation set has 1040 samples
BATCH_SIZE = 32
IMG_SHAPE = 224
EPOCHS = 20
def gen():
train_image_gen = ImageDataGenerator(rescale=1./255,
width_shift_range=0.1,
height_shift_range=0.1,
brightness_range=[0.2,1.0],
zoom_range=0.2,
horizontal_flip=True,
vertical_flip=True,
fill_mode='nearest')
train_gen = train_image_gen.flow_from_dataframe(train_data,
x_col='image_path',
y_col=[f'label_{x}' for x in np.arange(2)],
class_mode='raw',
batch_size=BATCH_SIZE,
shuffle=True,
target_size=(IMG_SHAPE,IMG_SHAPE))
val_image_gen = ImageDataGenerator(rescale=1./255)
val_gen = val_image_gen.flow_from_dataframe(val_data,
x_col='image_path',
y_col= [f'label_{x}' for x in np.arange(2)],
class_mode='raw',
batch_size=BATCH_SIZE,
target_size=(IMG_SHAPE,IMG_SHAPE))
return train_gen, val_gen
def plot(history):
training_accuracy = history.history['accuracy']
validation_accuracy = history.history['val_accuracy']
training_loss = history.history['loss']
validation_loss = history.history['val_loss']
epochs_range=range(len(training_accuracy))
plt.figure(figsize=(8, 8))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, training_accuracy, label='Training Accuracy')
plt.plot(epochs_range, validation_accuracy, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, training_loss, label='Training Loss')
plt.plot(epochs_range, validation_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
from PIL import Image
def predict(image_path, model):
im = cv2.imread(image_path)
test_image = np.asarray(im)
processed_test_image = process_image(test_image)
processed_test_image = np.expand_dims(processed_test_image, axis = 0)
ps = model.predict(processed_test_image)
return ps
def process_image(image):
image = tf.cast(image , tf.float32)
image = tf.image.resize(image , (224 , 224))
image = image/255
image = image.numpy()
return image
from tensorflow.keras.applications.vgg16 import VGG16
base = VGG16(weights = 'imagenet', include_top = False, input_shape = (224, 224, 3))
tf.keras.backend.clear_session()
for layer in base.layers:
layer.trainable = False
vgg_model = Sequential()
vgg_model.add(base)
vgg_model.add(GlobalAveragePooling2D())
vgg_model.add(BatchNormalization())
vgg_model.add(Dense(256, activation='relu'))
vgg_model.add(Dropout(0.5))
vgg_model.add(BatchNormalization())
vgg_model.add(Dense(128, activation='relu'))
vgg_model.add(Dropout(0.5))
vgg_model.add(Dense(2, activation='softmax'))
vgg_model.summary()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5 58892288/58889256 [==============================] - 0s 0us/step Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= vgg16 (Functional) (None, 7, 7, 512) 14714688 _________________________________________________________________ global_average_pooling2d (Gl (None, 512) 0 _________________________________________________________________ batch_normalization (BatchNo (None, 512) 2048 _________________________________________________________________ dense (Dense) (None, 256) 131328 _________________________________________________________________ dropout (Dropout) (None, 256) 0 _________________________________________________________________ batch_normalization_1 (Batch (None, 256) 1024 _________________________________________________________________ dense_1 (Dense) (None, 128) 32896 _________________________________________________________________ dropout_1 (Dropout) (None, 128) 0 _________________________________________________________________ dense_2 (Dense) (None, 2) 258 ================================================================= Total params: 14,882,242 Trainable params: 166,018 Non-trainable params: 14,716,224 _________________________________________________________________
train_gen, val_gen = gen()
optm = Adam(lr=0.0001)
vgg_model.compile(loss='binary_crossentropy', optimizer=optm,
metrics=['accuracy'])
EarlyStopping = EarlyStopping(monitor='val_loss',
min_delta=.0001,
patience=3,
verbose=1,
mode='auto',
restore_best_weights=True)
model_save = ModelCheckpoint('./vgg16_model.h5',
save_best_only = True,
save_weights_only = False,
monitor = 'val_loss',
mode = 'min', verbose = 1)
vgg_history = vgg_model.fit(train_gen,
steps_per_epoch = train_gen.samples // BATCH_SIZE,
epochs = 20,
validation_data = val_gen,
callbacks=[EarlyStopping, model_save],
class_weight = class_weights)
Found 4192 validated image filenames. Found 1040 validated image filenames. Epoch 1/20 131/131 [==============================] - 154s 1s/step - loss: 0.8175 - accuracy: 0.6022 - val_loss: 0.6654 - val_accuracy: 0.6798 Epoch 00001: val_loss improved from inf to 0.66541, saving model to ./vgg16_model.h5 Epoch 2/20 131/131 [==============================] - 96s 737ms/step - loss: 0.6610 - accuracy: 0.6953 - val_loss: 0.6565 - val_accuracy: 0.5587 Epoch 00002: val_loss improved from 0.66541 to 0.65652, saving model to ./vgg16_model.h5 Epoch 3/20 131/131 [==============================] - 97s 744ms/step - loss: 0.5702 - accuracy: 0.7364 - val_loss: 0.6720 - val_accuracy: 0.5452 Epoch 00003: val_loss did not improve from 0.65652 Epoch 4/20 131/131 [==============================] - 101s 770ms/step - loss: 0.5277 - accuracy: 0.7616 - val_loss: 0.5857 - val_accuracy: 0.6712 Epoch 00004: val_loss improved from 0.65652 to 0.58567, saving model to ./vgg16_model.h5 Epoch 5/20 131/131 [==============================] - 101s 773ms/step - loss: 0.4665 - accuracy: 0.8013 - val_loss: 0.5460 - val_accuracy: 0.7298 Epoch 00005: val_loss improved from 0.58567 to 0.54595, saving model to ./vgg16_model.h5 Epoch 6/20 131/131 [==============================] - 98s 750ms/step - loss: 0.4442 - accuracy: 0.8037 - val_loss: 0.4674 - val_accuracy: 0.7827 Epoch 00006: val_loss improved from 0.54595 to 0.46740, saving model to ./vgg16_model.h5 Epoch 7/20 131/131 [==============================] - 97s 743ms/step - loss: 0.4312 - accuracy: 0.8251 - val_loss: 0.4376 - val_accuracy: 0.8087 Epoch 00007: val_loss improved from 0.46740 to 0.43764, saving model to ./vgg16_model.h5 Epoch 8/20 131/131 [==============================] - 97s 740ms/step - loss: 0.3915 - accuracy: 0.8244 - val_loss: 0.3768 - val_accuracy: 0.8298 Epoch 00008: val_loss improved from 0.43764 to 0.37685, saving model to ./vgg16_model.h5 Epoch 9/20 131/131 [==============================] - 99s 753ms/step - loss: 0.3973 - accuracy: 0.8382 - val_loss: 0.4190 - val_accuracy: 0.8096 Epoch 00009: val_loss did not improve from 0.37685 Epoch 10/20 131/131 [==============================] - 101s 775ms/step - loss: 0.3613 - accuracy: 0.8405 - val_loss: 0.3667 - val_accuracy: 0.8452 Epoch 00010: val_loss improved from 0.37685 to 0.36667, saving model to ./vgg16_model.h5 Epoch 11/20 131/131 [==============================] - 98s 747ms/step - loss: 0.3427 - accuracy: 0.8565 - val_loss: 0.3431 - val_accuracy: 0.8548 Epoch 00011: val_loss improved from 0.36667 to 0.34311, saving model to ./vgg16_model.h5 Epoch 12/20 131/131 [==============================] - 98s 750ms/step - loss: 0.3791 - accuracy: 0.8448 - val_loss: 0.3543 - val_accuracy: 0.8471 Epoch 00012: val_loss did not improve from 0.34311 Epoch 13/20 131/131 [==============================] - 96s 737ms/step - loss: 0.3611 - accuracy: 0.8546 - val_loss: 0.3471 - val_accuracy: 0.8510 Epoch 00013: val_loss did not improve from 0.34311 Epoch 14/20 131/131 [==============================] - 97s 744ms/step - loss: 0.3182 - accuracy: 0.8738 - val_loss: 0.3377 - val_accuracy: 0.8529 Epoch 00014: val_loss improved from 0.34311 to 0.33768, saving model to ./vgg16_model.h5 Epoch 15/20 131/131 [==============================] - 96s 733ms/step - loss: 0.3369 - accuracy: 0.8561 - val_loss: 0.3083 - val_accuracy: 0.8596 Epoch 00015: val_loss improved from 0.33768 to 0.30827, saving model to ./vgg16_model.h5 Epoch 16/20 131/131 [==============================] - 97s 738ms/step - loss: 0.3379 - accuracy: 0.8638 - val_loss: 0.3041 - val_accuracy: 0.8673 Epoch 00016: val_loss improved from 0.30827 to 0.30413, saving model to ./vgg16_model.h5 Epoch 17/20 131/131 [==============================] - 96s 733ms/step - loss: 0.3522 - accuracy: 0.8608 - val_loss: 0.3468 - val_accuracy: 0.8500 Epoch 00017: val_loss did not improve from 0.30413 Epoch 18/20 131/131 [==============================] - 96s 734ms/step - loss: 0.3254 - accuracy: 0.8581 - val_loss: 0.2951 - val_accuracy: 0.8683 Epoch 00018: val_loss improved from 0.30413 to 0.29512, saving model to ./vgg16_model.h5 Epoch 19/20 131/131 [==============================] - 97s 736ms/step - loss: 0.3668 - accuracy: 0.8482 - val_loss: 0.3941 - val_accuracy: 0.8288 Epoch 00019: val_loss did not improve from 0.29512 Epoch 20/20 131/131 [==============================] - 96s 732ms/step - loss: 0.3221 - accuracy: 0.8632 - val_loss: 0.2957 - val_accuracy: 0.8712 Epoch 00020: val_loss did not improve from 0.29512
plot(vgg_history)
vgg_pred =[]
for image in test_data.image_path:
vgg_pred.append(predict(image , vgg_model))
final_vgg_pred = np.argmax(vgg_pred , axis=-1)
actual_label = test_data['label']
print(classification_report(actual_label, final_vgg_pred))
matrix=confusion_matrix(actual_label, final_vgg_pred)
sns.heatmap(matrix,square=True, annot=True, fmt='d', cbar=False,
xticklabels=['0', '1'],
yticklabels=['0', '1'])
plt.xlabel('Predicted label')
plt.ylabel('True label');
precision recall f1-score support
0 0.77 0.89 0.83 234
1 0.93 0.84 0.88 390
accuracy 0.86 624
macro avg 0.85 0.87 0.86 624
weighted avg 0.87 0.86 0.86 624
print(vgg_history.history['val_accuracy'][-3])
print(vgg_history.history['val_loss'][-3])
0.8682692050933838 0.2951194941997528
from tensorflow.keras.applications.mobilenet_v2 import MobileNetV2
base = MobileNetV2(weights = 'imagenet', include_top = False, input_shape = (224, 224, 3))
tf.keras.backend.clear_session()
for layer in base.layers:
layer.trainable = False
mobilenet_model = Sequential()
mobilenet_model.add(base)
mobilenet_model.add(GlobalAveragePooling2D())
mobilenet_model.add(BatchNormalization())
mobilenet_model.add(Dense(256, activation='relu'))
mobilenet_model.add(Dropout(0.5))
mobilenet_model.add(BatchNormalization())
mobilenet_model.add(Dense(128, activation='relu'))
mobilenet_model.add(Dropout(0.5))
mobilenet_model.add(Dense(2, activation='softmax'))
mobilenet_model.summary()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/mobilenet_v2/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_224_no_top.h5 9412608/9406464 [==============================] - 0s 0us/step Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= mobilenetv2_1.00_224 (Functi (None, 7, 7, 1280) 2257984 _________________________________________________________________ global_average_pooling2d (Gl (None, 1280) 0 _________________________________________________________________ batch_normalization (BatchNo (None, 1280) 5120 _________________________________________________________________ dense (Dense) (None, 256) 327936 _________________________________________________________________ dropout (Dropout) (None, 256) 0 _________________________________________________________________ batch_normalization_1 (Batch (None, 256) 1024 _________________________________________________________________ dense_1 (Dense) (None, 128) 32896 _________________________________________________________________ dropout_1 (Dropout) (None, 128) 0 _________________________________________________________________ dense_2 (Dense) (None, 2) 258 ================================================================= Total params: 2,625,218 Trainable params: 364,162 Non-trainable params: 2,261,056 _________________________________________________________________
train_gen, val_gen = gen()
optm = Adam(lr=0.0001)
mobilenet_model.compile(loss='binary_crossentropy', optimizer=optm,
metrics=['accuracy'])
EarlyStopping = EarlyStopping(monitor='val_loss',
min_delta=.0001,
patience=3,
verbose=1,
mode='auto',
restore_best_weights=True)
model_save = ModelCheckpoint('./mobilenetV2.h5',
save_best_only = True,
save_weights_only = False,
monitor = 'val_loss',
mode = 'min', verbose = 1)
mob_history = mobilenet_model.fit(train_gen,
steps_per_epoch = train_gen.samples // BATCH_SIZE,
epochs = EPOCHS,
validation_data = val_gen,
callbacks=[EarlyStopping, model_save])
Found 4192 validated image filenames. Found 1040 validated image filenames. Epoch 1/20 131/131 [==============================] - 98s 730ms/step - loss: 0.7897 - accuracy: 0.6004 - val_loss: 0.4541 - val_accuracy: 0.8500 Epoch 00001: val_loss improved from inf to 0.45408, saving model to ./mobilenetV2.h5 Epoch 2/20 131/131 [==============================] - 94s 720ms/step - loss: 0.4763 - accuracy: 0.8183 - val_loss: 0.3734 - val_accuracy: 0.8462 Epoch 00002: val_loss improved from 0.45408 to 0.37338, saving model to ./mobilenetV2.h5 Epoch 3/20 131/131 [==============================] - 94s 717ms/step - loss: 0.3754 - accuracy: 0.8612 - val_loss: 0.2983 - val_accuracy: 0.8760 Epoch 00003: val_loss improved from 0.37338 to 0.29825, saving model to ./mobilenetV2.h5 Epoch 4/20 131/131 [==============================] - 94s 719ms/step - loss: 0.3288 - accuracy: 0.8731 - val_loss: 0.2885 - val_accuracy: 0.8721 Epoch 00004: val_loss improved from 0.29825 to 0.28847, saving model to ./mobilenetV2.h5 Epoch 5/20 131/131 [==============================] - 94s 715ms/step - loss: 0.2855 - accuracy: 0.8919 - val_loss: 0.2511 - val_accuracy: 0.8971 Epoch 00005: val_loss improved from 0.28847 to 0.25111, saving model to ./mobilenetV2.h5 Epoch 6/20 131/131 [==============================] - 94s 718ms/step - loss: 0.2626 - accuracy: 0.8992 - val_loss: 0.2335 - val_accuracy: 0.9077 Epoch 00006: val_loss improved from 0.25111 to 0.23352, saving model to ./mobilenetV2.h5 Epoch 7/20 131/131 [==============================] - 94s 716ms/step - loss: 0.2499 - accuracy: 0.9072 - val_loss: 0.2273 - val_accuracy: 0.9058 Epoch 00007: val_loss improved from 0.23352 to 0.22725, saving model to ./mobilenetV2.h5 Epoch 8/20 131/131 [==============================] - 96s 730ms/step - loss: 0.2306 - accuracy: 0.9154 - val_loss: 0.2161 - val_accuracy: 0.9077 Epoch 00008: val_loss improved from 0.22725 to 0.21608, saving model to ./mobilenetV2.h5 Epoch 9/20 131/131 [==============================] - 93s 710ms/step - loss: 0.2201 - accuracy: 0.9248 - val_loss: 0.2136 - val_accuracy: 0.9096 Epoch 00009: val_loss improved from 0.21608 to 0.21362, saving model to ./mobilenetV2.h5 Epoch 10/20 131/131 [==============================] - 94s 716ms/step - loss: 0.2271 - accuracy: 0.9127 - val_loss: 0.1906 - val_accuracy: 0.9240 Epoch 00010: val_loss improved from 0.21362 to 0.19060, saving model to ./mobilenetV2.h5 Epoch 11/20 131/131 [==============================] - 94s 716ms/step - loss: 0.2344 - accuracy: 0.9072 - val_loss: 0.2223 - val_accuracy: 0.9077 Epoch 00011: val_loss did not improve from 0.19060 Epoch 12/20 131/131 [==============================] - 94s 715ms/step - loss: 0.2009 - accuracy: 0.9250 - val_loss: 0.2177 - val_accuracy: 0.9058 Epoch 00012: val_loss did not improve from 0.19060 Epoch 13/20 131/131 [==============================] - 94s 716ms/step - loss: 0.2195 - accuracy: 0.9127 - val_loss: 0.2025 - val_accuracy: 0.9173 Restoring model weights from the end of the best epoch. Epoch 00013: val_loss did not improve from 0.19060 Epoch 00013: early stopping
plot(mob_history)
mob_pred =[]
for image in test_data.image_path:
mob_pred.append(predict(image , mobilenet_model))
final_mob_pred = np.argmax(mob_pred , axis=-1)
actual_label = test_data['label']
print(classification_report(actual_label, final_mob_pred))
matrix=confusion_matrix(actual_label, final_mob_pred)
sns.heatmap(matrix,square=True, annot=True, fmt='d', cbar=False,
xticklabels=['0', '1'],
yticklabels=['0', '1'])
plt.xlabel('Predicted label')
plt.ylabel('True label');
precision recall f1-score support
0 0.88 0.85 0.86 234
1 0.91 0.93 0.92 390
accuracy 0.90 624
macro avg 0.89 0.89 0.89 624
weighted avg 0.90 0.90 0.90 624
print(mob_history.history['val_accuracy'][-4])
print(mob_history.history['val_loss'][-4])
0.9240384697914124 0.1905973255634308
from tensorflow.keras.applications.densenet import DenseNet169
from tensorflow.keras.applications.densenet import preprocess_input as densenet_preprocess
base = DenseNet169(weights = 'imagenet', include_top = False, input_shape = (224, 224, 3))
tf.keras.backend.clear_session()
for layer in base.layers:
layer.trainable = False
densenet_model = Sequential()
densenet_model.add(base)
densenet_model.add(GlobalAveragePooling2D())
densenet_model.add(BatchNormalization())
densenet_model.add(Dense(256, activation='relu'))
densenet_model.add(Dropout(0.5))
densenet_model.add(BatchNormalization())
densenet_model.add(Dense(128, activation='relu'))
densenet_model.add(Dropout(0.5))
densenet_model.add(Dense(2, activation='softmax'))
densenet_model.summary()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/densenet/densenet169_weights_tf_dim_ordering_tf_kernels_notop.h5 51879936/51877672 [==============================] - 0s 0us/step Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= densenet169 (Functional) (None, 7, 7, 1664) 12642880 _________________________________________________________________ global_average_pooling2d (Gl (None, 1664) 0 _________________________________________________________________ batch_normalization (BatchNo (None, 1664) 6656 _________________________________________________________________ dense (Dense) (None, 256) 426240 _________________________________________________________________ dropout (Dropout) (None, 256) 0 _________________________________________________________________ batch_normalization_1 (Batch (None, 256) 1024 _________________________________________________________________ dense_1 (Dense) (None, 128) 32896 _________________________________________________________________ dropout_1 (Dropout) (None, 128) 0 _________________________________________________________________ dense_2 (Dense) (None, 2) 258 ================================================================= Total params: 13,109,954 Trainable params: 463,234 Non-trainable params: 12,646,720 _________________________________________________________________
train_gen, val_gen = gen()
optm = Adam(lr=0.0001)
densenet_model.compile(loss='binary_crossentropy', optimizer=optm,
metrics=['accuracy'])
EarlyStopping = EarlyStopping(monitor='val_loss',
min_delta=.0001,
patience=3,
verbose=1,
mode='auto',
restore_best_weights=True)
model_save = ModelCheckpoint('./densenet169.h5',
save_best_only = True,
save_weights_only = False,
monitor = 'val_loss',
mode = 'min', verbose = 1)
dense_history = densenet_model.fit(train_gen,
steps_per_epoch = train_gen.samples // BATCH_SIZE,
epochs = EPOCHS,
validation_data = val_gen,
callbacks=[EarlyStopping, model_save])
Found 4192 validated image filenames. Found 1040 validated image filenames. Epoch 1/20 131/131 [==============================] - 109s 768ms/step - loss: 0.8379 - accuracy: 0.5811 - val_loss: 0.4822 - val_accuracy: 0.8596 Epoch 00001: val_loss improved from inf to 0.48221, saving model to ./densenet169.h5 Epoch 2/20 131/131 [==============================] - 99s 755ms/step - loss: 0.4852 - accuracy: 0.8235 - val_loss: 0.2789 - val_accuracy: 0.9337 Epoch 00002: val_loss improved from 0.48221 to 0.27895, saving model to ./densenet169.h5 Epoch 3/20 131/131 [==============================] - 98s 743ms/step - loss: 0.3651 - accuracy: 0.8727 - val_loss: 0.2052 - val_accuracy: 0.9452 Epoch 00003: val_loss improved from 0.27895 to 0.20515, saving model to ./densenet169.h5 Epoch 4/20 131/131 [==============================] - 97s 744ms/step - loss: 0.3211 - accuracy: 0.8859 - val_loss: 0.1630 - val_accuracy: 0.9529 Epoch 00004: val_loss improved from 0.20515 to 0.16301, saving model to ./densenet169.h5 Epoch 5/20 131/131 [==============================] - 97s 741ms/step - loss: 0.2703 - accuracy: 0.9098 - val_loss: 0.1440 - val_accuracy: 0.9567 Epoch 00005: val_loss improved from 0.16301 to 0.14397, saving model to ./densenet169.h5 Epoch 6/20 131/131 [==============================] - 96s 737ms/step - loss: 0.2594 - accuracy: 0.9120 - val_loss: 0.1354 - val_accuracy: 0.9577 Epoch 00006: val_loss improved from 0.14397 to 0.13536, saving model to ./densenet169.h5 Epoch 7/20 131/131 [==============================] - 97s 742ms/step - loss: 0.2400 - accuracy: 0.9163 - val_loss: 0.1398 - val_accuracy: 0.9548 Epoch 00007: val_loss did not improve from 0.13536 Epoch 8/20 131/131 [==============================] - 97s 737ms/step - loss: 0.2261 - accuracy: 0.9214 - val_loss: 0.1261 - val_accuracy: 0.9615 Epoch 00008: val_loss improved from 0.13536 to 0.12606, saving model to ./densenet169.h5 Epoch 9/20 131/131 [==============================] - 97s 738ms/step - loss: 0.2188 - accuracy: 0.9200 - val_loss: 0.1352 - val_accuracy: 0.9548 Epoch 00009: val_loss did not improve from 0.12606 Epoch 10/20 131/131 [==============================] - 97s 739ms/step - loss: 0.1961 - accuracy: 0.9271 - val_loss: 0.1310 - val_accuracy: 0.9587 Epoch 00010: val_loss did not improve from 0.12606 Epoch 11/20 131/131 [==============================] - 96s 735ms/step - loss: 0.2053 - accuracy: 0.9205 - val_loss: 0.1284 - val_accuracy: 0.9577 Restoring model weights from the end of the best epoch. Epoch 00011: val_loss did not improve from 0.12606 Epoch 00011: early stopping
plot(dense_history)
dense_pred =[]
for image in test_data.image_path:
dense_pred.append(predict(image , densenet_model))
final_dense_pred = np.argmax(dense_pred , axis=-1)
actual_label = test_data['label']
print(classification_report(actual_label, final_dense_pred))
matrix=confusion_matrix(actual_label, final_dense_pred)
sns.heatmap(matrix,square=True, annot=True, fmt='d', cbar=False,
xticklabels=['0', '1'],
yticklabels=['0', '1'])
plt.xlabel('Predicted label')
plt.ylabel('True label');
precision recall f1-score support
0 0.91 0.82 0.86 234
1 0.90 0.95 0.92 390
accuracy 0.90 624
macro avg 0.90 0.88 0.89 624
weighted avg 0.90 0.90 0.90 624
print(dense_history.history['val_accuracy'][-4])
print(dense_history.history['val_loss'][-4])
0.9615384340286255 0.12605918943881989
from tensorflow.keras.applications import InceptionV3
base = InceptionV3(weights = 'imagenet', include_top = False, input_shape = (224, 224, 3))
tf.keras.backend.clear_session()
for layer in base.layers:
layer.trainable = False
incept_model = Sequential()
incept_model.add(base)
incept_model.add(GlobalAveragePooling2D())
incept_model.add(BatchNormalization())
incept_model.add(Dense(256, activation='relu'))
incept_model.add(Dropout(0.5))
incept_model.add(BatchNormalization())
incept_model.add(Dense(128, activation='relu'))
incept_model.add(Dropout(0.5))
incept_model.add(Dense(2, activation='softmax'))
incept_model.summary()
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/inception_v3/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5 87916544/87910968 [==============================] - 1s 0us/step Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= inception_v3 (Functional) (None, 5, 5, 2048) 21802784 _________________________________________________________________ global_average_pooling2d (Gl (None, 2048) 0 _________________________________________________________________ batch_normalization (BatchNo (None, 2048) 8192 _________________________________________________________________ dense (Dense) (None, 256) 524544 _________________________________________________________________ dropout (Dropout) (None, 256) 0 _________________________________________________________________ batch_normalization_1 (Batch (None, 256) 1024 _________________________________________________________________ dense_1 (Dense) (None, 128) 32896 _________________________________________________________________ dropout_1 (Dropout) (None, 128) 0 _________________________________________________________________ dense_2 (Dense) (None, 2) 258 ================================================================= Total params: 22,369,698 Trainable params: 562,306 Non-trainable params: 21,807,392 _________________________________________________________________
train_gen, val_gen = gen()
optm = Adam(lr=0.0001)
incept_model.compile(loss='binary_crossentropy', optimizer=optm,
metrics=['accuracy'])
EarlyStopping = EarlyStopping(monitor='val_loss',
min_delta=.0001,
patience=3,
verbose=1,
mode='auto',
restore_best_weights=True)
model_save = ModelCheckpoint('./inceptionV3.h5',
save_best_only = True,
save_weights_only = False,
monitor = 'val_loss',
mode = 'min', verbose = 1)
incept_history = incept_model.fit(train_gen,
steps_per_epoch = train_gen.samples // BATCH_SIZE,
epochs = EPOCHS,
validation_data = val_gen,
callbacks=[EarlyStopping, model_save])
Found 4192 validated image filenames. Found 1040 validated image filenames. Epoch 1/20 131/131 [==============================] - 102s 741ms/step - loss: 0.8389 - accuracy: 0.5565 - val_loss: 0.4376 - val_accuracy: 0.8510 Epoch 00001: val_loss improved from inf to 0.43757, saving model to ./inceptionV3.h5 Epoch 2/20 131/131 [==============================] - 96s 730ms/step - loss: 0.4806 - accuracy: 0.8190 - val_loss: 0.3199 - val_accuracy: 0.8904 Epoch 00002: val_loss improved from 0.43757 to 0.31991, saving model to ./inceptionV3.h5 Epoch 3/20 131/131 [==============================] - 95s 729ms/step - loss: 0.3748 - accuracy: 0.8638 - val_loss: 0.2753 - val_accuracy: 0.9010 Epoch 00003: val_loss improved from 0.31991 to 0.27534, saving model to ./inceptionV3.h5 Epoch 4/20 131/131 [==============================] - 95s 728ms/step - loss: 0.3201 - accuracy: 0.8842 - val_loss: 0.2437 - val_accuracy: 0.9048 Epoch 00004: val_loss improved from 0.27534 to 0.24369, saving model to ./inceptionV3.h5 Epoch 5/20 131/131 [==============================] - 95s 726ms/step - loss: 0.3072 - accuracy: 0.8820 - val_loss: 0.2194 - val_accuracy: 0.9240 Epoch 00005: val_loss improved from 0.24369 to 0.21937, saving model to ./inceptionV3.h5 Epoch 6/20 131/131 [==============================] - 97s 741ms/step - loss: 0.2951 - accuracy: 0.8898 - val_loss: 0.2161 - val_accuracy: 0.9279 Epoch 00006: val_loss improved from 0.21937 to 0.21605, saving model to ./inceptionV3.h5 Epoch 7/20 131/131 [==============================] - 96s 730ms/step - loss: 0.2682 - accuracy: 0.8935 - val_loss: 0.2126 - val_accuracy: 0.9240 Epoch 00007: val_loss improved from 0.21605 to 0.21263, saving model to ./inceptionV3.h5 Epoch 8/20 131/131 [==============================] - 95s 723ms/step - loss: 0.2459 - accuracy: 0.9160 - val_loss: 0.2071 - val_accuracy: 0.9240 Epoch 00008: val_loss improved from 0.21263 to 0.20712, saving model to ./inceptionV3.h5 Epoch 9/20 131/131 [==============================] - 96s 729ms/step - loss: 0.2449 - accuracy: 0.9149 - val_loss: 0.1880 - val_accuracy: 0.9327 Epoch 00009: val_loss improved from 0.20712 to 0.18799, saving model to ./inceptionV3.h5 Epoch 10/20 131/131 [==============================] - 96s 733ms/step - loss: 0.2139 - accuracy: 0.9227 - val_loss: 0.1776 - val_accuracy: 0.9413 Epoch 00010: val_loss improved from 0.18799 to 0.17756, saving model to ./inceptionV3.h5 Epoch 11/20 131/131 [==============================] - 99s 757ms/step - loss: 0.2120 - accuracy: 0.9260 - val_loss: 0.1844 - val_accuracy: 0.9346 Epoch 00011: val_loss did not improve from 0.17756 Epoch 12/20 131/131 [==============================] - 95s 727ms/step - loss: 0.2379 - accuracy: 0.9061 - val_loss: 0.1803 - val_accuracy: 0.9394 Epoch 00012: val_loss did not improve from 0.17756 Epoch 13/20 131/131 [==============================] - 95s 723ms/step - loss: 0.2258 - accuracy: 0.9154 - val_loss: 0.1758 - val_accuracy: 0.9413 Epoch 00013: val_loss improved from 0.17756 to 0.17578, saving model to ./inceptionV3.h5 Epoch 14/20 131/131 [==============================] - 96s 732ms/step - loss: 0.2083 - accuracy: 0.9258 - val_loss: 0.1823 - val_accuracy: 0.9298 Epoch 00014: val_loss did not improve from 0.17578 Epoch 15/20 131/131 [==============================] - 95s 726ms/step - loss: 0.2107 - accuracy: 0.9238 - val_loss: 0.1937 - val_accuracy: 0.9250 Epoch 00015: val_loss did not improve from 0.17578 Epoch 16/20 131/131 [==============================] - 96s 730ms/step - loss: 0.1975 - accuracy: 0.9246 - val_loss: 0.1863 - val_accuracy: 0.9308 Restoring model weights from the end of the best epoch. Epoch 00016: val_loss did not improve from 0.17578 Epoch 00016: early stopping
plot(incept_history)
incept_pred =[]
for image in test_data.image_path:
incept_pred.append(predict(image , incept_model))
final_incept_pred = np.argmax(incept_pred , axis=-1)
actual_label = test_data['label']
print(classification_report(actual_label, final_incept_pred))
matrix=confusion_matrix(actual_label, final_incept_pred)
sns.heatmap(matrix,square=True, annot=True, fmt='d', cbar=False,
xticklabels=['0', '1'],
yticklabels=['0', '1'])
plt.xlabel('Predicted label')
plt.ylabel('True label');
precision recall f1-score support
0 0.90 0.64 0.75 234
1 0.81 0.96 0.88 390
accuracy 0.84 624
macro avg 0.86 0.80 0.81 624
weighted avg 0.85 0.84 0.83 624
